Neolabs that locked in GPUs early can resell them at 2-3x, and Modal's founder predicts 'neolab to neocloud pivots'
What happened? A founder of a "neolab", one of the new wave of AI research labs trying to train frontier models, has a fallback plan if the models don't work out: resell the compute. Databricks' Yuchen Jin, a co-founder and former CTO of GPU cloud Hyperbolic, posted on 24 October 2026 that the unnamed founder told him the lab locked in GPU capacity early and, with H100, B200 and B300 rental prices up 2-3x since, "we can just resell the compute and print money". Modal founder Erik Bernhardsson quote-posted it with a prediction: "We're going to see a bunch of neolab to neocloud pivots next year." The original post passed 780,000 views. The founder's claim is anecdotal and the lab is not named.
Why it's plausible GPUs have not lost value the way many expected. Andreessen Horowitz's Charts of the Week, citing rental-price data from Ornn, says every Nvidia GPU it tracks rents for more than on 1 January: about +160% for the H200, +120% for the B200, +65% for the RTX PRO 6000 and +35% for the H100, and even six-year-old A100s cost more. a16z's Moses Sternstein writes that current prices are 9% to 143% above a straight-line six-year depreciation path, while noting the indices are imperfect proxies. The Information reported in August that the compute crunch was hitting neolabs hardest, with one-year contracts "almost impossible to get" and providers pushing three- to five-year terms.
The other side Replies to Bernhardsson called the pivot "a rebrand with a GPU attached" and predicted consolidation. The Economist reported this month that stress is building among GPU landlords: CoreWeave's five-year bond yields about 15%, only two of ten listed neoclouds are profitable, and their combined debt is about nine times annual revenue. It also argues compute is becoming a traded commodity, citing BCG's estimate of $360 billion spent on AI compute last year, rising to $2.3 trillion of transactions by 2030, with exchanges and startups building rental-price indices and futures. If supply catches up, the resale margin that makes the neolab hedge work could shrink.
Sources: Yuchen Jin on X, Erik Bernhardsson on X, a16z Charts of the Week, The Information, The Economist, The Economist on AI debt
